Chapter 20 : Applying Recommender Systems for Learning Analytics : A Tutorial

main methods used in recommender systems; they recommend an item to the user by comparing the representation of the item’s content with the user’s like-minded users and introduce them as so-called nearest neighbours to some target user; then they predict an item’s rating for that user on the basis of the ratings given to this item by the target users’ nearest neighbours (co-ratings) (Herlocker, Konstan, Terveen,

With the emergence of massive amounts of data in various domains, recommender systems have become a practical approach to provide users with the most suitable information based on their past behaviour and and turn the abundance from a problem into an asset such as educational data mining, big data, and Web data.For instance, data mining approaches can make recommendations based on similarity patterns detected from the collected data of users.Furthermore, as an important part of LA research.
Recommender systems can be differentiated according to their underlying technology and algorithms.Roughly, they are either content-based or use collaborative main methods used in recommender systems; they recommend an item to the user by comparing the representation of the item's content with the user's like-minded users and introduce them as so-called nearest neighbours to some target user; then they predict an item's rating for that user on the basis of the ratings given to this item by the target users' nearest neighbours (co-ratings) (Herlocker, Konstan, Terveen, 2012;Schafer, Frankowski, Herlocker, & Sen, 2007).
In the past, we have applied recommender systems in various educational projects with different objectives regarding the development and evaluation of recommender system algorithms in education; we especially As described by the RecSysTEL working group for Recommender Systems in Technology-Enhanced 2015) it is important to apply a standard evaluation method.The working group identified a research methodology consisting of four critical steps for eval-uating a recommender system in education: 1.A selection of dataset(s) that suit the recommendation task.For instance, the recommendation items for a user.
2. An of different algorithms on the selected datasets including well-known datasets (if possible, education-oriented datasets such as MovieLens makes movie recommendations) to provide insights into the performance of the recommender systems.
3. A comprehensive user study to test psycho-educational effects on learners as well as on the technical aspects of the designed recommender system.
4. A deployment of the recommender system in a real life application, where it can be tested under realistic, normal operational conditions with actual users.
The above four steps should be accompanied by a complete description of the recommender system reported in the special section on educational datasets 1 and made available for other researchers under certain conditions allow other researchers to repeat and adjust any part of the research to gain comparable results and new insights and thus build up a body of knowledge around recommender systems in learning analytics.
imental study that followed the research methodology described above for recommender systems in of a recommender system study that followed the then, we conclude.
In this section, we describe how one should evaluate a recommender system in learning, making use of an methodology described above.To this methodology, however, we added an additional step: that of devel-2013)., which is presented in a RecSysTEL special issue 1 (Manouselis et al., 2012).
In our study, our target environment is social learning platforms in general.Social learning platforms work similarly to social networks such as Facebook but, the purpose of learning and knowledge sharing.They for educational stakeholders such as teachers, students, learners, policy makers, and so on.Our target social 2 As The interface has been designed with students, teachers, parents and policy makers in mind.
it will empower stakeholders through a single, integrated access point for eLearning resources from dispersed educational repositories.Secondly, it engages stakeholders in the production of meaningful educational activities by using a social-network style multilingual portal, offering eLearning resources as well as services for the production of educational activities.Thirdly, it will assess the impact of the new educational activities, which could serve as a prototype to be adopted by stakeholders in school education.
ommender system can best suit the data and information needs of a social learning platform, the main for users.In the following sub-sections, we describe the study step by step.

Dataset Selection
type of data.In our case, the target social learning We chose the MACE and OpenScout datasets for the following reasons: 1.The datasets provide social data of users (ratings, tags, reviews, et cetera) on learning resources.So, the structure, content, and target users of the 2. Running recommender algorithms on these datasets helps us to evaluate their performance before 3.Both the MACE and OpenScout datasets comply -2 http:/ /opendiscoveryspace.eu

A RECOMMENDER SYSTEM EXPERIMENT IN THE EDUCATIONAL DOMAIN
for storing social data.
Besides these two datasets, we also tested the Mov-ieLens dataset as a reference since, up until now, the educational domain has been lacking reference datasets for study, unlike the ACM RecSys conference series, which deals with recommender systems in general.Table 20.1 provides an overview of all three datasets sparsity.All the data are described more fully our

Off line Data Study
Algorithms.In this second step, we tried to select algorithms that would work well with our data.First, it is important to check the input data to be fed into data, thus the data of the selected datasets, includes interaction data of users with learning resources (items).2. We ran the model-based CFs, including statesample data.
3. algorithms from steps 1 and 2. In addition to the baselines, we evaluated a graph-based approach neighbours using the conventional k-nearest Performance Evaluation.After choosing suitable datasets and recommender algorithms, we arrive at the task of evaluating the performance of candidate protocol (Herlocker et al., 2004).A good description of an evaluation protocol should address the following questions: Q1.What is going to be measured?
we measure the prediction accuracy of the recommendations generated.By this, we want to measure how much the rating predictions differ from the actual ones by comparing a training set and a test set.The training and test sets result from splitting our user ratings data (the same as user interaction data).In our metrics range from 1 to 5.
If the input data contains implicit user preferences, such as views, bookmarks, downloads, et cetera, we of the F1 score since it combines precision and recall, which are both important metrics in evaluating the accuracy and coverage of the recommendations generated (Herlocker et al., 2004).F1 ranges from 0 to 1.
ommendations on which a metric is measured, also known as a cut-off the F1 for the top 10 recommendations of the result set for each user.shows the values of F1.As Figure 20.2 shows, the graph-based approach performs best for MACE (8%) and MovieLens (24%) and the selected memory-based and model-based CFs come in second and third place right after the graph-based CF.For OpenScout, the memory-based approach performs better with a difference of almost 1%.
In conclusion, according to the results presented in Figure 20.2, the graph-based approach seems to perby an improved F1, which is an effective combination of precision and recall of the recommendation made.

Deployment of the Recommender System and User Study
In the educational domain, the importance of user et al., 2015).Since the main aim of recommender systems in education goes beyond accurate predictions, usefulness, novelty, and diversity of the recommendations.However, the majority of recommender system probably because user studies are time consuming and complicated.
by conducting a user study with our target platform.For this, we integrated the algorithms that performed made for them.For this we used a short questionnaire versity, and serendipity.The full description and results of this data study and the follow-up user study have not been published yet.The user study does not conrun user studies that can go beyond the success indictors of data studies, such as prediction accuracy.Accuracy is one of the important metrics in evaluating recommender systems but relying solely on this metric can lead data scientists and educational technologists down less effective pathways.
Accessing most educational datasets is challenging since they are not publicly and openly available, for the same datasets, and some of the algorithms used differ from their results.Therefore, we could not gain additional information from the comparisons One possible reason is that the studies use different versions of the same dataset because the collected data belongs to different periods of time.For the MACE dataset, for instance, different versions are available.system community.This problem originates from the fact that, unfortunately, there is no gold-standard dataset in the educational domain comparable to the MovieLens dataset3 in the e-commerce world.In fact, the LA community is in need of several representative datasets that can be used as approaches.The main aim is to achieve a standard data format to run LA research.This idea was initially and later followed up by the SoLAR Foundation for this lack of comparable results and the pressing need for a research cycle that uses data repositories to project called LinkedUp4 follows a promising approach towards providing a set of gold-standard datasets by ers-Lee, 2009).The LinkedUp project aims to provide a linked data pool for learning analytics research and to run several data competitions through the central data pool.
Overall, the outcomes of different recommender sysdomain are still hardly comparable due to the diversity of algorithms, learner models, datasets, and evaluation The main goal of this chapter has been to illustrate how to identify the most appropriate recommender system for a learning environment.To do so, we folevaluating recommender systems in learning.The methodology consists of four main steps: 1. Select suitable datasets preferably from the educational domain and, in case the actual data is not available yet, similarly to the target data.
2. Run a set of candidate recommender algorithms step should reveal which recommender algorithms best works with the input data.
3. Conduct a user study to measure user satisfaction on the recommendations made for them.

target learning platform.
importance of running user studies even though they are quite time consuming and complicated.
not represent the opinions of the European Union, and the European Union is not responsible for any use that might be made of its content.The work of EU project LACE.

PRACTICAL IMPLICATIONS AND LIMITATIONS CONCLUSION
Therefore, we chose to use the Collaborative Filtering (CF) family of recommender systems.CF algorithms rely on the interaction data of users, such as ratings, bookmarks, views, likes, et cetera, rather than on the content data used by content-based recommenders.CF recommenders can be either memory-based or either item-based or user-based, referring to the our study, we made use of all types and techniques: both memory-based and model-based, as well as both user-based and item-based.Figure 20.1 shows our 1.We compared performance of memory-based CFs, including both user-based and item-based, by employing different similarity functions.

Finally
Figure 20.2 shows the F1 results of best performing

Figure 20 . 2 .
Figure 20.2.F1 of the graph-based CF and the best performing baseline memory-based and model-based CFs